Spiral-Net with F1-based Optimization For Image-Based Crack Detection

Takumi Kobayashi

Abstract: Detecting cracks on concrete surface images is a key inspection for maintaining infrastructures such as bridge and tunnels. From the viewpoint of computer vision, the task of automatic crack detection poses two challenges. First, since the cracks are visually depicted by subtle patterns and also exhibit similar appearance to the other structural patterns, it is difficult to discriminatively characterize such less distinctive and finer defects. Second, the cracks are scarcely found, making the number of training samples for cracks significantly smaller than that of the other normal samples to be distinguished from the cracks. This is regarded as a class imbalance problem where the classifier is highly biased toward majority classes. In this study, we propose two methods to address these issues in the framework of deep learning for crack detection: a novel network, called Spiral-Net, and an effective optimization method to train the network. The proposed network is extended from U-Net to extract more detailed visual features, and the optimization method is formulated based on F1 score (F-measure) for properly learning the network even on the highly imbalanced training samples. The experimental results on crack detection demonstrate that the two proposed methods contribute to performance improvement individually and jointly.